Portail HAL Ensta
Not a member yet
11080 research outputs found
Sort by
How does a social scientist become a deeptech entrepreneur? The strengths and limitations of the Paris-Saclay academic innovation ecosystem
No full textInternational audienc
Emergence of Collective Open-Ended Exploration from Decentralized Meta-Reinforcement Learning
No full textInternational audienceRecent works have proven that intricate cooperative behaviors can emerge in agents trained using meta reinforcement learning on open ended task distributions using self-play. While the results are impressive, we argue that self-play and other centralized training techniques do not accurately reflect how general collective exploration strategies emerge in the natural world: through decentralized training and over an open-ended distribution of tasks. In this work we therefore investigate the emergence of collective exploration strategies, where several agents meta-learn independent recurrent policies on an open ended distribution of tasks. To this end we introduce a novel environment with an open ended procedurally generated task space which dynamically combines multiple subtasks sampled from five diverse task types to form a vast distribution of task trees. We show that decentralized agents trained in our environment exhibit strong generalization abilities when confronted with novel objects at test time. Additionally, despite never being forced to cooperate during training the agents learn collective exploration strategies which allow them to solve novel tasks never encountered during training. We further find that the agents learned collective exploration strategies extend to an open ended task setting, allowing them to solve task trees of twice the depth compared to the ones seen during training. Our open source code as well as videos of the agents can be found on our companion website
Heated press welding: analysis of the parameters influencing the mechanical strength of hybrid PA66/PA12 thermoplastic and S235 steel sheet joints
No full textInternational audienceThis study aims to optimize and characterize hot pressed polyamide/steel hybrid joints as a potential solution for reducing energy costs in public transport through structural lightweighting. The mechanical performance of these joints is compared to traditional bonding methods, specifically investigating two types of joints: PA66/steel and PA12/steel. A design of experiments approach is employed to explore the influence of process parameters, with a particular focus on the cooling rate and heating temperature as key factors. By determining the optimal process parameters, the study emphasizes the importance of reaching the polymer pyrolysis point and achieving the fastest cooling rate to achieve optimal results. The findings reveal that these hybrid joints exhibit comparable average shear strength values to bonded joints, showcasing their potential as effective joining methods. In conclusion, future developments in hybrid polymer/metal joining processes utilizing thermal methods should prioritize rapid and uniform heating at the polymer/metal interface to initiate pyrolysis selectively at the polymer surface for bond formation. Subsequent rapid cooling is essential to cease pyrolysis and prevent polymer degradation within its volume. These insights are crucial for successful implementation of such processes in various industrial applications
Thermomechanical shape memory testing of 4D printed novel material rhombus-shape structure
No full textInternational audience4D printing of functional energy generation/absorption structures by material extrusion technique can capitalize on the exciting applications in intelligent damping devices and patterns to deform spontaneously. This paper investigated the capitalization of this acquired knowledge by studying the shape memory effects of a functional rhombus shape structure. Initially, the shape memory characteristics of energy absorption and dissipation capacity were analyzed using dynamic mechanical testing to develop databases to gather thermophysical data before expressing the material behavior and recovery performances of printed shape memory structures. Then a complete thermomechanical cycle (shape programming and recovery) on 4D printed amorphous and semi-crystalline shape memory polymers exhibited deep insights into shape memory performance. The performance factors (i.e., recovery and fixity ratios) are influenced by variable printing parameters (i.e., layer height, printing temperature, and speed) and stimuli-based testing conditions. Results reveal both materials have significant shape memory effects with a maximum recovery ratio of up to 92.30 ± 0.32% from the programmed configuration. The amorphous polymer was extremely affected by printing temperature, whilst the semi-crystalline was influenced heavily by the interaction of all three parameters. Finally, shape memory effects predicted by a high-order design model and compared with experimental results showed negligible error. The analyses and assessments presented in this paper are adequate to understand the shape memory behavior under process control parameters to establish a data-driven model of a 4D printed semi-crystalline and amorphous polymer reactive to thermal stimuli
Evaluation expérimentale de l'hybridation des composites en fibre carbone unidirectionnelles à matrice polymère : comportement mécanique sous chargements quasistatiques et dynamiques
No full textIn the context of the increased use of carbon fiber-reinforced composites in the aerospace industry, this thesis investigates the influence of inter-ply hybridization, involving variations in carbon fiber type and ply thickness, on the internal structure, mechanical properties, and dynamic responses of unidirectional carbon fiber-reinforced composites. Focused on aerospace applications, the study utilizes two types of unidirectional carbon fiber prepregs with distinct fiber types and ply thicknesses, manufacturing both reference and hybrid laminates. The research objectives encompass comprehensive material characterization under quasi-static and dynamic loading conditions. Experimental analyses include monotonic on and off-axis quasi-static tests, off-axis cyclic load-unload tensile tests, coupled damage-plasticity model development, shockwave propagation assessment, dynamic tensile strength evaluation, and examination of damage position and threshold variations. Beyond contributing to the fundamental understanding of unidirectional composites, this research introduces insights into ply-level hybridization, offering potential applications in complex structures. As the aerospace industry demands materials with enhanced mechanical performance, this thesis provides valuable insights for advancing composite material design and application.Dans le contexte de l'utilisation croissante des composites renforcés à fibres de carbone dans l'industrie aérospatiale, cette thèse examine l'influence de l'hybridation entre plis, impliquant des variations de type de fibre de carbone et d'épaisseur de pli, sur la structure interne, les propriétés mécaniques et les réponses dynamiques des composites renforcés à fibres de carbone unidirectionnelles. Axée sur les applications aérospatiales, l'étude utilise deux types de préimprégnés de fibres de carbone unidirectionnelles avec des types de fibres et des épaisseurs de pli distincts, permettant la fabrication de laminés de référence et hybrides. Les objectifs de la recherche englobent une caractérisation complète des matériaux dans des conditions de chargement quasi-statique et dynamique. Les analyses expérimentales comprennent des essais quasi-statiques monotones en traction selon les axes et hors axes, des essais de traction cycliques hors axes avec chargement et déchargement, le développement d'un modèle couplé de dommage-plasticité, une évaluation de la propagation des ondes de choc, une évaluation de la résistance dynamique à la traction, et l'examen des localisations de dommage et des variations de seuil d’efforts à la rupture. Au-delà de contribuer à la compréhension fondamentale des composites unidirectionnels, cette recherche apporte des éclairages sur l'hybridation au niveau des plis, offrant des applications potentielles dans des structures complexes. Alors que l'industrie aérospatiale demande des matériaux avec des performances mécaniques améliorées, cette thèse fournit des perspectives précieuses pour faire progresser la conception et l'application des matériaux composites
Investigating the Impact of Atmospheric Boundary Layer Stratification on Wind Farm Noise Propagation
No full textInternational audienceThe expansion of wind farm installations has been hindered by annoyance issues resulting from the noise emitted by wind turbines. Understanding the factors that affect sound propagation is crucial to mitigate the impact of noise. Atmospheric boundary layer (ABL) stratification strongly affects the noise propagation of isolated wind turbines. However, few studies have looked at the influence of atmospheric conditions on wind farm noise propagation. This study aims to investigate this through numerical simulations. Large eddy simulations (LES) are used to predict the mean flow inside and around the wind farm. The noise from each wind turbine is computed from an extended source model that determines the wind turbine sound production based on its geometry, and on the flow characteristics (wind speed and turbulence intensity). A model based on the parabolic equation is employed to compute the sound propagation based on the flow fields obtained from LES. Neutral, stable and unstable stability conditions are considered for an idealized wind farm layout. The results of this study provide insight into the influence of atmospheric conditions on wind farm sound</div
Réduction de l’oubli catastrophique à l’aide de méthodes de distillation et de transfert de caractéristiques pour l’apprentissage incremental profond
No full textDeep learning methods are designed for closed-set recognition, where a predefined set of known classes is assumed. However, in real-world scenarios, open-set recognition is more realistic, allowing for the possibility of encountering unknown or novel classes during testing. Classincremental learning specifically addresses this problem by focusing on continuously improving models through the incorporation of new categories over time. Unfortunately, deep neural networks trained in this manner suffer from catastrophic forgetting, resulting in significant performance degradation on previously encountered classes. While various methods have been explored to alleviate this issue by rectifying theclassification layer of deep incremental models, an equally important aspect resides in the degradation of feature quality, which can impede downstream classification performance. This thesis specifically focuses on investigating diverse approaches to enhance feature quality, facilitating adaptation to new classes while preserving discriminative features for past classes during incremental training. Specifically two methods have been proposed and rigorously evaluated on widely established benchmarks, attaining performances either comparable or superior to the state-of-the-art . The first approach presented investigates the use of contrastive methods during incremental learning in order to improve the features extracted by incremental models while the second one uses an expansion and compression scheme to greatly reduce the forgetting happening at a feature level.Les méthodes d’apprentissage profond actuelles sont conçues de manière statique, avec un certain nombre de classes à reconnaître connu et prédéfini à l’avance. L’apprentissage incrémental, en revanche, étudie le scénario plus réaliste dans lequel certaines classes inconnues peuvent arriver au fur et à mesure et le modèle doit s’adapter et apprendre à reconnaître ces nouvelles catégories. Malheureusement, les réseaux neuronaux profonds entrainés de cette manière souffrent d’un oubli catastrophique, provoquant une dégradation significative des performances sur les classes précédemment rencontrées. Diverses méthodes ont été explorées pour atténuer ce problème en rectifiant la couche de classification des modèles incrémentaux. Cependant, un aspect tout aussi important réside dans la dégradation de la qualité des caractéristiques que le modèle extrait des images. Cette thèse se concentre spécifiquement sur l’exploration de diverses approches visant à améliorer leur pouvoir discriminant, facilitant ainsi l’adaptation à de nouvelles classes tout en préservant les caractéristiques discriminatives des classes précédentes lors de l’entraînement incrémental. Plus précisément, deux méthodes ont été proposées et rigoureusement évaluées, atteignant des performances comparables où supérieures à l’état de l’art. La première approche présentée explore l’utilisation de méthodes contrastives pendant l’apprentissage incrémental afin d’améliorer les caractéristiques extraites par le modèle, tandis que la seconde utilise une stratégies d’expansion et de compression de la partie responsable de l’extraction des caractéristiques dans le réseau de neurone afin de réduire significativement l'oubli
Dynamics of electronic states in the insulating intermediate surface phase of 1T−TaS
No full textInternational audienceThis article reports a comparative study of bulk and surface properties in the transition metal dichalcogenide 1T−TaS. When heating the sample, the surface displays an intermediate insulating phase that persists for ∼10 K on top of a metallic bulk. The weaker screening of Coulomb repulsion and a stiffer charge density wave (CDW) explain such resilience of a correlated insulator in the topmost layers. Both time-resolved angle-resolved photoelectron spectroscopy and transient reflectivity are employed to investigate the dynamics of electrons and CDW collective motion. It follows that the amplitude mode is always stiffer at the surface and displays variable coupling to the Mott-Peierls band, stronger in the low-temperature phase and weaker in the intermediate one
Robust water diffusion modeling in a structural polymer joint based on experimental X-ray tomographic data at the micrometer scale
No full textInternational audienceStructural bonding is a technique increasingly used in the industrial field. For applications in aggressive environments such as seawater, predicting the effect of moisture on the mechanical behavior of bonded assemblies is of paramount importance. The objective of this work is to analyze water diffusion in an epoxy adhesive material and, more specifically, to propose a robust method for choosing the most appropriate diffusion model. Experimental studies of the water absorption in a two-component epoxy structural adhesive, using gravimetry and X-ray tomography, were first performed. The presence of a population of pore-type defects in the polymeric joint helped to characterize the evolution of water diffusion kinetics. Thus, two diffusion mechanisms were identified: a first one related to the migration of water molecules within the adhesive matrix, and a second one related to the penetration of water into the pores. Then, Dual-Fick and Langmuir models were retained, as the two diffusion models most likely to capture the above mechanisms. Although it was shown that both models could give similar results in terms of global diffusion behavior, the results arising from these two models differ at the local scale, especially for extended periods of time. Therefore, special attention was paid to the second absorption mechanism, and a comparison of waterfronts between theoretical predictions and experimental tomographic data was achieved, leading to the final choice of a Dual-Fick diffusion model
Generating Constraint Programs for Variability Model Reasoning: A DSL and Solver Agnostic Approach
No full textInternational audienceVerifying and configuring large Software Product Lines (SPL) requires automation tools. Current state-of-the-art approaches involve translating Variability Models (VM) into a formalism accepted as input by a constraint solver. There are currently no standards for the Variability Modeling Languages (VML). There is also a variety of constraint solver input languages. This has resulted in a multiplication of ad-hoc architectures and tools specialized for a single pair of VML and solver, fragmenting the SPL community. To overcome this limitation, we propose a novel architecture based on model-driven code generation, where the syntax and semantics of VMLs can be declaratively specified as data, and a standard, humanreadable, formal pivot language is used between the VML and the solver input language. This architecture is the first to be fully generic by being agnostic to both VML and the solver paradigm. To validate the genericity of the approach, we have implemented a prototype tool together with declarative specifications for the syntax and semantics of two different VMLs and two different solver Families. One VML is for classic, static SPL (Feature Model) and the other is for run-time reconfigurable dynamic SPL with soft constraints to be optimized during configuration. The two solver families are Constraint Satisfaction Program (CSP) and Constraint Logic Programming (CLP). CCS Concepts: • Software and its engineering → Software product lines; Software architectures; • Computing methodologies → Model verification and validation